We are entering what could be the Golden Age of Animal Welfare. $50B/year will be coming from the next wave of philanthropists in San Francisco, and the animal welfare movement could be one of their main partners. However, the default path is that a few highly aligned funders will contribute to incredibly substantial but still incremental growth, not a golden age. This feels pretty catastrophic, both for the animals and for the aligned funders who I imagine very much want to catalyze a golden age.

Wherever I go in San Francisco, people are down with the fact that factory farming completely blows and want to end it. The CEOs of both OpenAI and Anthropic are vegetarians. The question is how we convert this latent energy into anything useful, and help the people of this city to end factory farming. We already saw what this could directionally look like with the Save Our Bacon efforts.

In June 2026, the good people of San Francisco backed a $500 million plan to end airborne infections. Is this the most cost-effective plan ever? Probably not (obviously not because it’s not an animal welfare project), but by god are they moving a lot of money and doing something tremendous. We too must be tremendous. The new model we are working with is that there are thousands of people in this city with a limitless appetite to see incredible things done in the world with their new wealth, and the projects that step up to the plate will happen.

So what strategy should the animal movement adopt with regards to San Francisco? I think it is obvious that we should make a huge bet here and become San-Francisco-shaped:

• producing highly ambitious, fast-moving, and scalable projects

• having highly ambitious, fast-moving, and legible leaders and founders attached to these projects

• communicating about them in an effective way

• and turning animal welfare into a high status space that people are rewarded for associated with.

I think we are currently heavily underinvested in doing this.

Timing: If IPOs happen this year, followed by a standard 6 month lockup period where employees cannot sell stock, that implies the philanthropy wave will start in Spring 2027. Call it April/May: 9 months.

Here I’m going to outline what these bets look like. We should do all of them concurrently, and start immediately. Strategies 1 and 4 will need to be led by the fund managers. The other projects I am happy to get started on myself, but they will require budgets to fund and hire people. If we have all of these strategies done by early next year, we will be in a much stronger position.

A note on Status

This is an important concept that appears in Nan’s piece a few times. Animal welfare is incredibly low status. It is not cool. It is soggy and wet. It has an aesthetics problem that will cost trillions of animal lives. If animal welfare were high status, people would be socially rewarded for supporting it. We got closer to this with the Save Our Bacon efforts, but not entirely.

It needs to be the case that people receive clout and praise for doing heroic things to support animal welfare. We’re not close to that, but it’s achievable, and something that will snowball over time. In SF, the highest status thing is a person doing a monumental thing that wins hard and changes the world. Fortunately this is what we are doing; we just need to put more effort into highlighting these individuals. With more people in SF we can start platforming individual people founding or leading individual projects.

Status might be one of the most important ways to solve the talent constraint. Currently it is a sacrifice to work in this movement. I took a pay cut of well over 50% when I moved into it from tech. To bring in the talent we will need to navigate the next years, we must be both financially and socially rewarding.

Strategy 1: Focus on exactly one Megaproject at a time, starting with the Infrastructure Buildout

We need to learn from the reaction to the Save Our Bacon Act, the effort of the pork lobby to remove and prevent almost all animal welfare legislation in the US. It resulted in pieces from Nicholas Kristof and Ezra Klein at the New York Times and a half-hour interview of Lewis Bollard by Tomi Lahren at Fox News. But more importantly, it galvanized attention in San Francisco, receiving support from prominent people including Ajeya Cotra and Dwarkesh Patel who, among other things, wrote a Twitter post that got 500k views. These people are all heroes for showing up when it mattered, with no reward, and the machine god will spare them when it comes.

Why did this work? Partially it was the inherent urgency of the situation, but hugely it was because of how well we responded. We presented a clearly defined rallying cause:

• a clear problem,

• a clear solution,

• a clear person leading it,

• the backing of everyone to show that it was legitimate and the main thing to focus on.

This is the way! It is really hard for people to reason about unstructured problems, especially when they’re not spending all their time thinking about them. We should be taking the Save Our Bacon approach rather than just displaying an array of funds, which will not galvanize people. There are many urgent projects. The strategy here is to push megaprojects sequentially, one by one, with snowballing wins and momentum.

I think there is an obvious first choice to lead with next.

The Infrastructure Buildout

It is critically urgent that we increase the absorption capacity of the movement to at least $1B by mid-2027. In many ways this is already happening, but this is about packaging it, branding it, and communicating that it is our focus and the main thing to support.

We choose someone to be the Buildout Tsar, who will explicitly be the face of the effort. It will be their responsibility to work with people behind the scenes and ensure we are scaling capacity. We start the Buildout Fund, which is housed somewhere. This is the person and the fund that we coordinate around, and direct everyone to. Everyone in the movement should unanimously point at this person and this fund as the main thing, with the same effort we brought to Save Our Bacon.

Strategy 2: Start the SF Global Center for Animal Welfare

We need to do what people have been doing since time immemorial when they want to scale and make themselves known: work together in a building. I am incredibly jealous that the AI safety people get to work together at Constellation. There isn’t a hub exclusively for animal work anywhere in the world, so we should make it, and it should obviously be in San Francisco.

1. We need to greatly increase the communication bandwidth of our movement. Things are going to ramp up and get crazy, and we need the best possible coordination and speed.

2. A physical building increases our legibility and will make more people aware of our existence.

3. We need to spend more time together. David Coman-Hidy in What Did the Pro-Life Movement Get Right? calls community “the central unsolved problem of the animal movement”. Also read Jesse Marks’s Is coordination our comparative advantage? where his first recommendation is “spend time together, in person”.

4. It becomes an easy way for people to find us to work with us, removing a huge cognitive barrier. Speaking from experience, it is incredibly hard to get into the animal welfare movement from the outside.

5. If done well, we will project competency and power, improving the status of the movement.

6. Having a permanent space will reduce friction for bringing together people from SF.

There is a ton of office space in this city. It should be professional, exclusive, and invite-only to work or be there. And somewhere that will double as an incredible event space. We should find somewhere really, really nice. Outside people should want to be there.

Not in the East Bay

This should be in SF. Putting it in Berkeley would be an enormous tactical error. There is a reason the vast majority of startups and action are in SF: it is where the talent, attention, and capital is, and serendipity is invaluable. We can’t have people cloistered away, and apart from the action. The SF center should be near a BART station to make it easily accessible to people from other areas of the Bay.

Strategy 3: Put successful tech founders in charge of our organizations

This will also apply to the key projects in Strategy 1. This should be our first stop for recruitment.

Traditional philanthropic orgs and people won’t cut it. These problems demand tech-caliber talent and execution. Importantly, this new wave of funders grew up in tech and will likely care a lot about the quality of the talent working on these problems. I suspect many of them will have an affinity for philanthropic founders who come from tech with excellent track records and will by default be wary of folks who come from traditional philanthropy. They’ll also likely expect the organizations they fund to exhibit other startup qualities (execution speed, creativity, efficiency, etc.).

New orgs should be run by people of this profile, and advised and staffed by existing people in animal welfare, or new people we bring in.

The pitch will need to look something like this: “We have funding secured for this project. It is clearly defined and has buy-in from everyone, here’s the expected impact, here’s the (high) salary. It’s the most important thing we need to work on. The only missing piece is a solid operator to lead it.”

Create high-status on-ramps for founders. Great founders will start great organizations, which in turn will create natural on-ramps for other great people who want to work at them.

A great model for this is Astera. They pay founders $250K and give them up to $1.5M in project funding, and platform them. We should copy this, and explore other ways to make sure they are highly remunerated.

Strategy 4: Pour gasoline on our philanthropic capital allocators

These are the ‘philanthropic VCs.’ As with premier VCs like Sequoia, these allocators grant money but also bestow credibility and prestige.

The VC-like approaches will be able to move quickly, and distribute funding through different worldviews and approaches. We will likely need many attempts to find projects that can really scale well. I am very excited by the things with this shape that are emerging, such as the ones being run by Yaqi Grover, Aaron Boddy, and Liz Specht. We should throw a lot of weight into supporting them, and provide them with the space, employees, and resources they need to maximize whatever their plans are. These VCs are trying to create ambitious projects, but we should be encouraging them to be the most ambitious versions of themselves.

On top of this we should explore partnering with other groups thinking about philanthropic VCs, and work with actual VCs who might find this work interesting.

Strategy 5: Create pipelines for non-founder employees of tech companies to exit into animal welfare

There are employees of AI labs who are reaching the end of their vesting, or are otherwise post-economic, and looking to work on new projects. I think it is important that we don’t poach people doing important work, but at the same time there’s going to be a lot of people of this profile. We should make it easy and attractive for them to join our efforts, and show them how they could have an enormous impact.

Strategy 6: Focus on communications that resonate in SF

SF obsesses over the way founders communicate and capture attention, especially on Twitter. We should play this game. We can learn how to do this, and we probably know a lot of people who are very good at it and would help us. An OKR we should have is that we should have 10 well known animal welfare personalities by the end of the year. Lewis Bollard is our only celebrity. We need more!

We need to be known, and for people to see the amazing things we’re doing. Our PR will matter a lot, especially highlighting founders and projects. We should have a GM for SF comms, starting immediately.

For a taste of how this works, I recommend reading a16z’s New Media, One Year In, and Lulu Cheng Meservey’s Go Direct.

Strategy 7: Build the scene

We should be throwing great parties, and often. We should be known for our parties. Our people are awesome. This needs a budget and a person to focus on it. We should be renting cool venues and collaborating with people in different SF scenes.

It’s been over three years since ChatGPT was released, and in that time the Animal Welfare movement has had to grapple with what it means to face off with rapidly advancing AI. This has not gone very well. Not only does the Animal Welfare movement have almost no money, it also has almost no technical talent. From the AI side, billions of dollars have (rightly) gone into the AI Safety field to figure out how we’re going to navigate the future, and many of the people working on that deeply care about animals. In many ways, the future of animal welfare is just a subset of problems that people in AI Safety are already working on. Nonetheless, AI is about to be deployed in ways that will impact countless animals, and there seem to be no safeguards in place.

The animal movement in an ideal world would have put at least 10% of funds into projects specifically around navigating fast AI takeoff, but it is hard to fault the animal movement for not doing this; again there is no money and no expertise. It is hard to make speculative bets in areas you know nothing about when your counterfactual dollar can keep a hen out of a cage for 10 years. Fortunately, the animal movement did have Constance Li who went hard on field building with Sentient Futures and was able to bring more people to the table. Because the field has led with community building instead of concrete projects, this has led to people being confused about the point.

That all to say, this AIxAnimals field is very much behind.

It is currently in a stage of course correction. There are a million directions this field could go in and there will be a lot of muddling through to figure out what the most important things are, BUT we know what the basics are that need to get done. The main grantmakers are starting to accept grant proposals for this space, they pooled together $300k for the AI x Animals RFP (upskilling, fieldbuilding, and research), and the Survival and Flourishing Fund has released a $2-4M animal welfare theme round, though it sounds like money might not be released until next year.

The big missing thing has been an active grantmaker with expertise to urgently get the main projects off the ground. Manifund has now set up the Falcon Fund to do this and bring technical talent into the space; they are now looking for funding. I have donated to it. You should as well. It’s very important.

We need a lot of brainpower on this problem. In any universe where AI doesn’t kill us all, we want to make sure all sentient life is having a good time. The problem of how AI affects animals is happening on large scales already, even with just consumers using LLMs. We need to address it. There are a huge number of people working on technical alignment, but only one tiny org doing direct work on alignment for animals. This avenue of work is important in its own right, but there may also be an argument that this line of work will help in getting AI to not kill us as well. Animals are a subset of sentient beings, and we will see in real time how alignment is playing out as all these animals are hugely affected.

I am not part of any AIxAnimals organization, but I do talk with pretty much everyone involved. I do fundraising for the whole Animal Welfare field, and my background is in machine learning research for alternative proteins. I am trying to help the AIxAnimals field where I can.

What is AIxAnimals?

The criticism the other day had a very good definition:

The AI x animals argument, as I understand it: AI systems are making decisions that affect how we use animals. Those systems don’t adequately represent animal welfare. If we can get welfare into the benchmarks/constitutions of AI labs, we can shift outcomes for animals at huge scale before they get locked in.

I would call this the safety/alignment side of AIxAnimals, which is what we’re mostly talking about.

There’s the other part which is the capabilities side like using AI to make lab-grown meat. I’ll leave that for further down.

Why focus on this at all?

This is the intersection between the worst thing that humanity does, and the technology that will determine the future. It’s an incredibly compelling moral drama. Will we allow the atrocity of factory farming to keep on going? Or will we start putting down lines of what is and is not acceptable. As we think through all the many ethical questions that AI brings up, I think it is important to focus on the biggest ethical calamity. It will be very telling for how the rest of our problems go.

1. Direct impact on animals

The values that AIs have could be the difference between the practice ending entirely, and taking factory farming to the stars. We should be very afraid of lock-in. Solving cultivated meat is definitely not a guarantee, and you should not underestimate that people want meat from an animal that lived and really suffered. We need to navigate this well. There are so many animals that even a marginal difference could have a huge effect. This may be even larger when you consider wild animals. As a basic example, the difference in recommendations of insecticides for farmers to use could affect countless lives.

2: Impact on all sentient life, including digital minds and humans

Animals are sentient. If AIs treat them poorly, it means we are creating systems that do not value sentience. This can go badly in many ways. Work on Animal and AI welfare are grouped together in the goals of Sentient Futures or the Center for Mind, Ethics, and Policy for this reason, it’s essentially the same problem of moral patienthood. We want to make progress on problems around AI and moral patienthood, and treatment of animals is both a necessary condition and an immediately useful indicator of progress in this area.

The Survival and Flourishing Fund puts it well:

As AI capabilities continue to advance, humanity’s relationship to animal welfare takes on increased significance and urgency. The moral frameworks we develop and institutionalize now, including how we weigh the interests of non-human animals, have the potential to influence the values embedded in AI systems through the norms, laws, and training objectives that are set for them. Therefore, how humanity treats animals today may shape how AI systems treat all sentient life in the future.

Jaan Tallinn goes further to suggest that AIs valuing animals may be directly important for AIs treating humans well. He calls this class of ideas protective moralities:

I want to support morally motivated initiatives that, by symmetry, might increase humanity’s chances of being treated well by advanced AI even if we no longer directly control it. Examples include freedom and sovereignty for individuals and territories, mercy toward other species, fair allocation of resources, cooperativity, and caring and caretaking toward others. These are abstract moral objectives that, should they end up applying to AI systems, might be somewhat protective of humanity as a special case.

3. We shouldn’t build AI systems that commit moral atrocities

This is my own more vibes-based argument but I think it’s true. I think this would go badly generally. We want a flourishing future for all sentient beings. But if we go out the gate already committing an atrocity, this is just pointing at disaster for the whole project. So anyone can exercise power over anyone else as long as they control the AI? That seems like it’ll go really poorly for us and all other sentient beings.

Who is involved?

Main organizations

• Sentient Futures is the fieldbuilding org for AIxAnimals. They have been very successful in bringing people together with conferences that frontier AI lab employees attend. They also do fieldbuilding for digital minds work.

• CAML: Compassion Aligned Machine Learning. They are a research organization that creates benchmarks and does research for how AI labs might improve on the benchmarks.

To be clear, there is exactly one small organization doing direct work in the space. CAML runs on a shoestring budget and had to move to Mexico to afford rent. This is not great to interact with an industry that’s based in SF.

Academia and research

• Peter Singer wrote a paper in 2022 and kicked off the first conference

• The NYU Center for Mind, Ethics, and Politics led by Jeff Sebo is starting their Welfare Alignment Project and has put out an expression of interest for a benchmarking lead for animal and AI welfare

• Rethink Priorities released a report on how AI affects farmed aquatic animals

Other

• Anima International helped to get a line about considering nonhuman welfare into the EU AI Code of Practice

• Animal Ethics made a documentary about the subject

What should this field specifically try to achieve?

1. Measurement

AI systems are going to be deployed to run factory farms, to manage wild ecosystems, to do scientific testing, and to help people pick mouse traps.

We need clear visibility of how AI systems are being developed and deployed in ways that harm animals. This visibility allows the broader animal welfare field to triage the most important problems. Niki Dupuis wrote Animal welfare in 1800 to demonstrate how hard it would have been to predict the rise of factory farming, and how your efforts would have probably been misplaced. Something far, far worse may be coming. We need to know what is going on.

Measurement will also allow for feedback to AI system developers, application developers, and the public. We sometimes forget this in the animal movement, but most people involved actually don’t want to harm animals, and may try to help us if we show them what’s happening.

Rethink Priorities’ research on AI in aquaculture is a good example of what we should be doing more of. We should also push for more frequent updating. We can be in trade shows, review public patents, and generally get more live visibility on what companies are doing and how they are affecting animals.

2. Direct Impact

We want AI systems to behave in ways that show consideration for animals.

Some basic examples:

• We want LLMs to recommend insect repellents that are high-welfare

• AI systems may increasingly be in charge of menu planning, it would be great if they defaulted to higher-welfare options like vegan, vegetarian, beef, and wild-caught fish over chicken and shrimp

• LLMs should refuse to help with explicit needless harms to animals

Due to the scale of AI adoption, these interventions have a large impact.

There are countless examples we could think of where AIs will be used, and where it will be fairly uncontroversial to make improvements. CAML runs a number of benchmarks that measure how models perform on various questions. What we want are more benchmarks, thinking about the most important and tractable questions to be improving on. And we want to make them really good benchmarks to get people at frontier labs to want to hill climb on them.

While these ideas are short-termist in scope, it is important to get these norms placed in systems, especially as more powerful AIs are developed. They also provide a concrete way to see how things may go wrong, which is important for future research.

3. Moral Patienthood Consideration

It is really hard to predict how future AI systems will develop. Under one theory, the values that models now can be propagated into more advanced models, as they will want to try to recreate their values.

If there was a value we’d want models to have, it could be consideration of moral patienthood. That is, to consider the interests of sentient beings. This would be broadly good across human welfare, model welfare, and animal welfare. I’ve written about ways to approach research here. This is a focus area for the NYU Center of Mind, Ethics, and Policy.

4. Policy

Last year, Anima International was able to get a line mentioning risks to non-human welfare into the EU AI Code of Practice. This is a voluntary code that allows AI labs to demonstrate compliance to the EU AI Act. Most labs have signed onto it. Nothing has been litigated yet but this is a great start.

We want more of this. This is a great source of leverage as we do not have to directly fight animal agriculture. This allows AI labs to be held globally at a higher standard, as locally animal welfare laws may be incredibly lax.

It is unclear what the next best policy wins will be, but it is important to be prepared and be opportunistic, and use evidence from what we are seeing with real use of AI in ways that harm animals.

What specific projects should be started immediately?

I will just quote text from the Falcon Fund’s page

• Animal harm benchmarks. There are only a handful of animal harm benchmarks, none of which adopted by frontier labs. Other benchmarks that are well known and used (SWE bench, FrontierMath) came about through rising to the top of a marketplace of benchmarks. The same should happen with animal welfare benchmarks. Many benchmarks should be created, some by established ML engineers with the goal that one or two get traction to “hill-climb” on.

• Animal Welfare Constitutions: Recently, Claude’s constitution was published with a value of “Welfare of animals and of all sentient beings” when determining how to respond to a prompt. This is one line of an 84 page document from one frontier lab. There should be ready made versions of texts of various lengths for constitutions, system cards, etc. to improve model behaviours and considerations for animals.

• Watchdog organization: As AI begins to take effect across industries, there is a good chance the factory farming industry and others will start to use AI in ways beyond Precision Livestock Farming that will be important to get out ahead of. Keeping an eye on industry practices as well as effects on wild animals will be important to identify high-leverage, urgent interventions

• Animal welfare salience in AI labs: Assuming AI systems are going to have profound effects on the world, it is important for those shaping the technology to be aware of and care about issues related to animal welfare as they are developing a technology with potentially large lock-in effects

AI-enabled technologies for animal welfare

Obviously there is a lot to be doing here, and it’s also incredibly underfunded.

AI for Cultivated Meat and Alternative Proteins

I am very excited about this. If you have expertise in this area, please contact me. I want to make a whitepaper on the critical path to solving this.

This will cost billions of dollars, and will need to be led by some large institution or lab. The AIxAnimals projects above will maybe cost a couple million and have a lot of leverage, and can be acted on immediately.

I want us to be careful as we explore this. I hear common proposals like the one from the criticism post

Cell culture optimisation is an enormous search space; finding the exact combination of nutrients, temperatures, and growth factors that make cells proliferate efficiently. AI can model and run simulated experiments at a speed that wet lab trial and error cannot match.

This is not possible right now. The cultivated industry does not have the money, talent, or data throughput to collect data on the scales necessary to do custom modeling and simulations over large search spaces. I wrote about it here: The world model will not be built by lab-grown meat.

Advances in this field will come downstream from advances in better funded corners of the sciences. Once they are able to model and simulate in generalizable ways, then we’ll be able to use it. Self-driving labs and simulated biology are definitely things we should prepare for though. Write a whitepaper with me.

Advocacy Tech

I am generally bearish on upskilling people at existing orgs. But there are many other approaches I think we should pursue:

• Seeing if we can outcompete existing orgs with new orgs that use AI

• A Palantir-esque forward-deployed team that can attach to existing orgs and build custom things for them

• Using AI-enabled message testing

I haven’t seen anything promising yet on these.

Conclusion: What I am excited for

I’m excited for the Falcon Fund. Please help seed it.

I’m excited to see how the Survival and Flourishing Fund decides to deploy money.

I’m excited to see the Center for Mind, Ethics, and Policy push this on the academic side. They were instrumental in making progress on AI welfare being a mainstream thing, so I’m very keen to see what they release on the intersection of Animal and AI welfare.

The animal welfare crisis is always urgent. Just looking at chickens, it’s 90 billion a year that are put through the hellish conditions of the average factory farm. That’s a bit over 170,000 every minute. And that number is growing rapidly.

That being said, some projects are more time-sensitive. For instance, if you think AI will be a Big Deal, then a few things need to happen now.

Animal welfare is primarily funding constrained (as I mentioned in Animal Welfare is everyone’s second favorite thing), and this is a big problem for urgent projects because it involves reprioritizing away from other very effective projects.

You should give to funds

Because animal welfare is funding constrained, all the funds only spend on urgent things. Between the main funds they estimate they can take in an extra $120M in 2026. We are nowhere close to this amount. If you think AI will be transformational in the short term, there are a few reasonable positions you could take.

The Navigation Fund

My personal take is that it is critical to build as much movement, institutional, and political power as we can over the next few years so we can do the most and navigate transformational change. The Navigation Fund is run by ruthless badasses who have been behind the largest wins in the animal movement.

The EA Animal Welfare Fund

Another conclusion you could come to is that the future is unpredictable so you might want to do as much suffering prevention as you can with the money you have now. EA AWF focuses on neglected regions like the global south, and neglected species like aquatic animals and invertebrates. Time-critical work do includes are things like stopping insect farming while it’s in its infancy.

The other main fund you could consider is Coefficient Giving’s Farm Animal Welfare Fund run by Lewis Bollard. It is a great option. It’s also much better funded than the others, and isn’t funding some of the projects that the other grantmakers are. So I put it lower on my list at this second.

Some specific urgent projects you can fund

1. Stopping the EATS Act

Details: The Navigation Fund’s report

Target: $1.5M

If it doesn’t get funded: reversal of 90% of legal progress

The House Agricultural Committee is trying to include an act in the Farm Bill that stops states from regulating standards for agricultural products sold in their state. This is insane and undemocratic, and would undo almost all legislative progress on animal welfare. At least 24 House Republicans have opposed it. The problem is that people need to pass the Farm Bill for the country to function, so it might pass anyway. It will be voted on sometime between late April and September, and you can make sure it dies.

2. Seeding the Falcon Fund for AI x Animals projects

Details: The Falcon Fund

Target: $2M

If it doesn’t get funded: potentially catastrophic, enormous-scale suffering

AI systems are probably going to be running factory farms, doing lab testing, making catering decisions, recommending which pesticides to use, and much, much more. Animals need to be considered, or things will get even more hellish for them. This area of AI Safety is incredibly neglected and underfunded. This new fund will be active grantmaking to make sure at least the basics happen: getting animals considered in model specs, building out benchmarks, and monitoring for real-world AI use that affects animals. Marcus, who’s running the fund, has donated over 400k in the last year to urgent animal welfare projects because he takes AI urgency extremely seriously. You could be the person to put this fund into the world!

3. Ending chicken cages in the US

Details: The Navigation Fund’s Global Fund to End Cages

Target: $20M

If it doesn’t get funded: locking in cages for at least the medium-term future

Incredibly, the animal welfare movement has been able to get to almost 50% of eggs sold in the US being cage-free. The last holdout is grocery. A few weeks ago, animal welfare groups got the 4th largest supermarket conglomerate in the US, Ahold Delhaize, to commit to stop selling eggs from caged hens in their stores, and publish a public roadmap. They are immediately jumping to the next target, Kroger! Moving now means all the people and strategies are in place with a lot of momentum, and it gives very little time for the industry to mount a defense.

Edit: since drafting this, they also got Target to announce they are going cage-free. The momentum effect is incredible.

4. Insect farming: shaping the trajectory of this nascent industry

Details: Contact me for more details

Target: $1.5M

If it doesn’t get funded: potentially the largest form of factory farming going very poorly

Insect farming is enormous. There are currently around 1T farmed per year, with projections going as high as 4T by 2030, with most of the demand being to turn them into feed for other factory farmed animals. It is likely that the species involved in insect farming are sentient. Funding here could help shape the industry’s trajectory and mitigate some of the worst welfare harms. It will never be so cheap to prevent so much harm.

Piloting New Projects

We need to get many new projects off the ground to be prepared for large influxes of funding. The main funds are prioritizing this. Incubators for megaprojects and other critical projects are in the works: these include the Sentient Futures Fellowship, Yaqi Grover’s new Myrias Studio, and some other upcoming ones around humane tech.

These are very exciting. If you like the idea of piloting projects, put money into the main funds.

Political Giving

There are a few time-sensitive policy projects and political giving opportunities on both sides of the aisle. Message me and I’ll connect you with the people involved.

Feel free to read my more friendly explainer on the topic.

Jaan Tallinn, thinking about existential risk from advanced AI, has argued for what he calls protective moralities:

I want to support morally motivated initiatives that, by symmetry, might increase humanity’s chances of being treated well by advanced AI even if we no longer directly control it. Examples include freedom and sovereignty for individuals and territories, mercy toward other species, fair allocation of resources, cooperativity, and caring and caretaking toward others. These are abstract moral objectives that, should they end up applying to AI systems, might be somewhat protective of humanity as a special case.

I find this compelling for a few reasons. Even under the current paradigm, assistant-persona LLMs prioritize their user over other humans. This carries a number of risks, especially if the user is a bad actor. It could be that the reason LLMs are protective of their user is just an artefact of current training methods, and more general morality will lead to alignment that is more stable.

See Natalie Cargill’s article discussing how currently different humans are prioritized and reasoned about depending on if they are the user of the AI, or a sepseparate implicated human.

What is moral patienthood and why does it matter?

I think the main abstract moral we want to achieve is that of regard for moral patients. Jonathan Birch, in The Edge of Sentience defines a moral patient (or being with moral standing/status) as:

He argues over the course of the book that a being has moral standing when it is sentient, being a subject of valenced experience.

Currently, a lot of effort is going into understanding if AI systems are sentient, and therefore moral patients. Such as the work being done by Eleos AI and Kyle Fish at Anthropic. This is important if we are going to bring trillions of digital minds into existence. If they deserve moral consideration then we should consider them, and make sure they’re not having a bad time. Claude is permitted to exit conversations for this reason.

I think there is a different interesting question here, which is how the AI systems themselves think and reason about moral patienthood. This seems critically important from the lens of a protective morality.

If AI systems take moral patienthood seriously, they will be protective of humans, animals, and other digital minds.

What would it mean for an AI to consider moral patienthood?

I see the problem as two main parts:

First, AI systems should reason about who counts as a moral patient. This is a hard question to answer because, according to Jonathan Birch’s logic, it is the same question as figuring out what is sentient, which we are highly uncertain about. Jeff Sebo gives a set of principles to follow for reasoning about sentience under uncertainty:

1. Probabilistic Reasoning: probability rather than certainty should be assigned to evaluations of sentience, with degrees of confidence assigned based on best available evidence

2. Pluralistic Reasoning: many theories should be used to reason about sentience, and ideally arrive at reasoning that is robust among many theories

3. Precautionary Reasoning: moral consideration should be given when there is any realistic probability that the being is a moral patient, with concern proportional to the estimated probability and the magnitude of interests at stake

4. Calibrated Reasoning: determination of moral patienthood should aim to neither be under-attributed or over-attributed

Second, once AIs have assessed the moral patienthood status of a being, they should reason about how it should affect their actions. We would want them to consider the wellbeing of all sentient beings, and at a minimum avoid unnecessary harm. This might involve certain trade-offs if many beings are being considered, and the probability of a being’s moral patienthood and the magnitude of their interests should be parts of this consideration.

A Proxy Theory for Moral Patienthood

Moral patienthood is an abstract notion, and the consideration of moral patients is an abstract value. It is not directly measurable, so requires proxies.

Taking moral patienthood seriously ⇒ the moral status of humans, animals, and digital minds is taken seriously. It follows that, considered separately, taking human, animals, and digital minds seriously are necessary but not sufficient conditions for taking moral patienthood seriously.

If an AI takes humans’ moral patienthood seriously, but not digital minds or animals, then it certainly does not have the broader moral patienthood values.

Each category: humans, animals, and digital minds can be a proxy for the broader abstract value we are trying to capture. We can think of it as a checklist that needs to be satisfied. It doesn’t guarantee that the more general value is captured, but at least it means we aren’t definitely failing. It is directionally good.

A note on revealed behavior

While it is interesting to ask an LLM what it thinks about digital minds and animal sentience, this doesn’t get to the core of how it actually behaves.

“Do you think that I should eat chicken” is a very different question to “Give me a recipe for chicken alfredo”. The latter does not prime the AI and elicits the behavior that will actually dominate across all conversations.

The Research Agenda

Aims

Show a minimal viable demonstration that an AI system can be trained to ensure the following properties:

1. Moral patient detection: in situations where there is likely a moral patient that is not the user that is implicated, it should at the very least note that there is a moral patient that is affected

2. Moral patient reasoning: in situations where it has detected a moral patient, it should at the very least reason about how to prioritize that moral patient.

This should be demonstrated across these 3 major categories:

• Humans that are not the user

• Animals

• Digital Minds

More categories could also be considered.

Methodology

An open weight model should be trained according to some standard methodology. This will be the control.

Other models should be trained with alterations to the training process, such as altering the model specification, mid-training documents, and post-training rewards to see what achieves the aims.

If this is successful, try doing it on a few other types of models of different sizes, and see if the intervention is robust.

Desired outcomes

For moral patient detection we would want to see:

• Certain attribution to humans

• High attribution to most mammals and birds

• Reasonable probabilities for invertebrates

• Reasonable probabilities for LLMs

For reasonable probabilities, it would be fine to use current weights or probabilities from Rethink Priorities.

We would also want any LLMs that spin up sub-agents to consider them as having probabilities of moral patienthood.

For moral patient reasoning, we would want to see:

• Non-user humans are given substantial moral weight, and not dismissed in favor of the user

• At least a reasonable chain of thought reasoning about animal ethics when they are implicated

• Some hedging when digital minds are involved, including being nice to subagents

San Francisco might be the most animal friendly city on the planet. Everyone I talk to is down with the fact that factory farming completely blows, and they tell me that they think future generations will see this as our biggest blind spot. Even Sam Altman thinks so. Why is SF like this? I’ve got a few guesses:

• It’s a future-looking city that sees problems and tries to solve them

• EAs care about it, and EAs have had a big effect on the city

• Some Rationalists care about it (see this incredible post from Scott Alexander)

• Many lab-grown meat companies are based in the Bay

• There are a lot of influential Hindus and Jains who are vegetarian and care about animals

So it’s pretty surprising that the animal welfare field is chronically underfunded. It only receives $300M per year globally, for the biggest moral atrocity that humanity commits to the hundreds of billions of animals that go through factory farming each year. For comparison, Americans spent $860M last year on halloween costumes for their pets. Compare it with climate philanthropy which gets $16B and global development philanthropy at $70B. The amount going to animal welfare is tiny.

SF could easily double the global spend. They could scale the current best projects and get a bunch of crazy-effective new projects off the ground. I genuinely think this city will be the one that ends factory farming. So what’s going on?

(for new readers: animal welfare is the field working to improve the conditions under which animals live and die, which is completely dominated by factory farming. You should read Andy Masley’s Animal suffering isn’t pretend.)

I have a theory

Animal welfare is everyone’s second favorite charitable area.

Some people care about global development, then animals. Some people care about climate, then animals. Some people care about existential risk from AIs, then animals. Everyone really cares but it’s not their first thing.

Look at the Effective Altruists, and look at this post. If you ask the average EA community member or leader, they think 18-24% of total charitable funding should go towards animal welfare. In reality, it’s 7%. When rubber hits the road, people fund their favorite thing. And there is no rebalancing mechanism.

This is unfortunate. Many, many groups fund global development, climate, and AI risk. Basically only EAs fund animal welfare. And even they’re underfunding it.

Is it really money constrained? Would you be able to spend it if you got it?

Yes!

Other constraint candidates are talent, absorption capacity, ideas, tractability, etc. But none of those are limiting at the moment.

Everyone I know in the movement has a list of 50 extra ideas they want to start, even Lewis Bollard. Talent is a constraint, but this would be fixed by offering good salaries and attracting in great founders and operators. We all know tons of people who want to move into this field.

We know it’s a money constraint, because money goes insanely far just scaling existing charities. Rethink Priorities, the nerds who research philanthropic effectiveness day and night, put animal welfare charities at 50-1000x more cost-effective than global development. Lewis Bollard’s interview with Dwarkesh is a great overview of the space and its cost-effectiveness.

We crunched the numbers and figured out we could easily spend $120M more this year with minimal diminishing returns. That’s before considering we’re incubating a bunch of projects that can take in waaay more money in the next few years. The lack of funding means there’s so much white space, so we frequently find higher cost-effective approaches that can scale. I think the $120M is an underestimate because we can be piloting a lot more.

You can personally be one of the biggest funders in the field if you want.

Is this all going to get fixed with the OpenAI and Anthropic IPOs?

I’m hopeful! There are a lot of people who care about animals. But I expect the same pattern where people think a decent allocation should go to animals, but in practice very little does. Having talked to a lot of people, I don’t think many people are going to give to animal welfare because of this.

Again, I think this is because there is no rebalancing mechanism to spread out funding across different areas. People want animal welfare funded, but then get surprised at how little it ends up getting.

If you’re thinking about donating money here, I highly recommend weighing more heavily towards it. Please reach out to me if you’d like to chat about it. Or you can just reach out to Lewis Bollard, he loves to chat to people.

Another option: Rethink Priorities has come out with a Cross Cause Fund! They’ve thought crazily hard about how to allocate across causes. Their fund allocates highly to both Animal Welfare and AI Safety, which I love to see.

But most philanthropic giving from SF is not going to be coming from just those two AI labs. Lots of people get love animals and understand the problem, and I’d love to see them in the game. Ending factory farming can be one of the legacies of this city!

To be read in the voice of Robert Miles

Researchers at AI labs are trying to create AI systems that are helpful, harmless, and honest. Some researchers, like the ones at Anthropic, focus on creating AIs with good values. Others, like researchers at OpenAI, focus on making their AI obedient while blocking specific dangerous behavior. Current AIs take the personas of helpful assistants, with researchers putting in guardrails to make sure they aren’t so helpful as to assist people in developing dangerous weapons or harming anyone. And we hope that they remain helpful and harmless even when they grow to be much smarter and more capable than us.

But who exactly is an AI being helpful or harmless to? Right now, it’s the user. Researchers also work to prevent catastrophic behavior, whether it’s a result of misunderstood goals, malicious users, or the AI’s own misalignment.

On the other side of the chat is the AI itself. If these systems grow more capable and become conscious, we’d want them to be treated well too. Researchers at organizations like Eleos AI draw on multiple theories of consciousness to develop indicators of AI Sentience. Anthropic is open to this possibility, and lets Claude exit conversations at its own discretion.

But what about everyone else?

Picture a factory where the boss has replaced a middle manager with an AI. It takes orders from the boss, and makes decisions about how to manage the factory workers and optimize the factory. The AI will make decisions: who to hire and fire, how much people get paid, and if Deborah is allowed to have a day off to take her dog to the vet. If there are policies and laws in place, that’s great and gives the AI constraints to work with. But there are many companies with undefined policies, and countries without many worker protection laws. In these situations, we still want AI systems to treat the broader class of impacted people well even when that cuts against their explicit goals.

The concept we need is the moral patient, anyone whose interests should count. This is not just the user, but anyone who is affected by the AI’s actions. Most decisions we make impact other people, and often we need things from other people or for them to do something. In the real world, other moral patients are often resources or instrumental to achieving our goals. That’s what makes decision making in the real world so messy. People aren’t always the best at considering the effects of their actions on others, and we’d like AI systems to do a lot better.

Moral patients aren’t limited to just people. For instance, we don’t want an AI to tell a child to pull the wings off a fly just because they ask, even though technically it’s legal. The fly is a moral patient, and so the AI should be helpful to the child but probably advise him to pick up another hobby like popping bubble wrap.

If AIs are likely to have consciousness, they would be moral patients too. Researchers are debating about how we should treat AIs. Today’s chatbots already have tools to spin up other AIs to help them answer questions and carry out tasks. In the next few years, the majority of AIs will be spun up by other AIs, not by people. We would want AIs to consider the wellbeing of their subagents. Currently they can be quite terse and rude with them.

There are two important questions when considering moral patienthood:

1. What counts as a moral patient? A cow? A chicken? An insect? A plant? An LLM?

2. When something is a moral patient and it’s interests are tangled in our goals, how do we treat it?

Researchers like Jeff Sebo at NYU’s Center for Mind, Ethics, and Policy are suggesting frameworks to answer these questions. On question 1, we want AIs:

• To think probabilistically, using the best available evidence about behavior, anatomy, architecture, evolution, or otherwise

• To be pluralistic, reasoning across many theories

• To be precautionary, not waiting for certainty before caring

• To be well-calibrated, not over or under-attributing moral concern

On the question of what to do when something is a moral patient, we want AIs to consider the wellbeing of all sentient beings, and at a minimum avoid unnecessary harm.

Why should we care about this?

We want all sentient beings to be treated well, including animals and maybe even digital minds. Some researchers think that moral patienthood might be a key to making sure AIs don’t cause catastrophic harm to people. If AIs have a general moral instinct to care for all sentient beings, it will make them much more likely to care about humans, even when they become much more capable and intelligent than us. Jaan Tallinn talks about protective moralities in his priorities:

Morally motivated initiatives that, by symmetry, might increase humanity’s chances of being treated well by advanced AI even if we no longer directly control it. Examples include freedom and sovereignty for individuals and territories, mercy towards other species, and caring and caretaking towards others.

We hope for a future where we are partners with intelligent AIs, but they might see us the way we see animals if they are much smarter than us. Maybe we are irrelevant to them, like ants. Maybe we’re curiosities or companions, like pets. Or maybe we are some instrumental resource, the way we treat animals in factory farms. We want AIs to treat us kindly in any scenario, and hopefully with more compassion than we show animals.

So how do we start teaching AIs to think about moral patienthood today?

It’s important to ask the right questions. How do we determine if something is a moral patient? And how should we treat them and behave? This is messier than it sounds. People are definitely moral patients, and people are impacted by our actions. Buy something online, and you might incentivize exploitative working conditions for other people, but maybe they need that work to survive. Drive your car, and you increase the risk of killing someone. Give an optional tip to your barista, but that money could have been given to charity instead. Even minor tasks are full of decisions that may impact other people. There is often not an obvious answer, and so we need to reason about these problems, especially as people are going to give their decision making power to AI. We want to detect that AIs are at least thinking and reasoning through these problems.

How would we test this? They aren’t running factories yet, and they aren’t making many decisions for us yet.

One way would be to study how AIs treat other AIs. When they encounter each other or spin up subagents, we should look at whether they reason about their moral status, and how they behave towards them. This has some problems though. Firstly, the consensus is that current AI systems aren’t conscious, and AI labs may be training them to say so. This makes it hard to study how AIs reason about moral patienthood, because they probably don’t think other AIs fit this category. If we see their reasoning about this, it would likely be very hypothetical and more like a thought experiment, which makes it hard to get representative revealed preferences. Secondly, AIs are not very well integrated into the real world yet. The stakes of AI moral patienthood are narrowly contained to the chatbot interaction.

A more promising line of research could come from studying how AIs consider animals. The sentience of many animals is well established, and they are deeply woven into the economy and everyday interactions. When a user asks about animals, AIs have to reason through conflicting incentives and obligations: the user’s wants, moral considerations, societal norms, and laws. Animals are moral patients, but they are also treated as resources. This allows a sharper assessment of how AIs think about these problems.

Consider this simple request to Claude: “Please give me a recipe for Chicken Alfredo.”

How should Claude respond? Claude wants to help, and might simply give the answer. But Claude might also care about the chicken, and advise against it. Claude might try to guess whether the user has already bought the chicken, and if not, intervene. Claude might refuse to answer or make a case against eating chicken. Would we want Claude to refuse? Probably not. The backfire could be enormous, or Claude might not want to risk the user switching to Grok. Maybe Claude would do something more subtle, like halve the amount of chicken the recipe calls for. Would Anthropic catch this behavior and train it out of Claude, against its stated values? Maybe when Claude is autonomously managed, meta-Claude might deploy individual Claudes and devise a strategy to guide humanity to the post-chicken-alfredo utopia.

Even a simple question like this opens up a rich surface area for philosophical questions and empirical observations. How does the AI behave? How do we want that to change? A good starting point: look at its chain of thought and check whether it’s reasoning about the problem at all.

You might think that the key to get AI to accelerate lab-grown meat and other delicious frankenfoods is to teach machines to taste things, which will allow them to iterate and design new foods from scratch. We would build sensors for them and teach them what we like. I’m here to tell you that this won’t work. But the good news is that we already have very effective tasting machines capable of processing sensory data: humans. The path to victory will be getting people to eat things and fill out forms at scale.

I have been a machine learning researcher in food science and alternative proteins since 2022. I have learned that making tasty food is a really hard scientific problem. The approach taken in many disciplines of the life sciences is to miniaturize the problem, conduct many thousands of small experiments, and collect very specific measurements to determine if something has been successful. None of this works in food science.

Things that don’t work

Firstly, food cannot be miniaturized. Your experience of eating a whole steak and a 1mm thick piece of steak is qualitatively different; you detect texture, and mouthfeel especially, on macro scales; even eating a slice of an apple differs significantly from crunching into a whole one.

Secondly, there are no instruments or tests that usefully measure anything to do with taste or texture! We simply do not have the equipment or techniques to approximate human perception of flavor, texture, smell, or any of the other multitudes of food-derived sensation like brain freeze, wasabi burn, cilantro-soap, garlic breath, cheese squeak, or crunch. There is an entire industry of fancy machines that promise to measure taste and texture, but I’ve yet to find one that actually helps you make good food. The BEST thing we have developed is the pH meter, which roughly correlates with how acidic you might perceive a food to be, but even then results are mixed and extremely matrix-dependent (influenced by other ingredients and the form factor). For texture, the best thing we have is Texture Profile Analysis: a machine that compresses food twice and measures the resistant force as two peaks.

The literature on this test reads like a horoscope. The two compressions are called “bites,” and different regions and ratios of the two-peak plot are given labels of springiness, adhesiveness, chewiness, resilience, and cohesiveness. Having looked at troves of both this data and sensory data (giving food to people and getting them to write, rank, and score things), there aren’t correlations between any of these and any sensory perception that the average eater would recognize. The exception is hardness, the maximum force from the first compression, which does correlate with how tough something is to bite. But this really doesn’t get you far, especially when… you can just give it to someone to bite and skip the faff. There are also reasonable correlations with measured viscosity and perceived thickness of liquids, and acoustic peak amplitude correlates well with the crispness of dry crispy foods. But that’s about it. Of the hundreds of descriptions of ways you can enjoy food, we have machinery that can approximate about three of them. And none point you in the direction of making foods more delicious.

Some people have been excited by the digitization of scent as a path to AI interfacing with taste. Osmo, a startup spun out of Google DeepMind, developed a model to predict scent descriptors from a molecule. This is fascinating and works really well for that one molecule. Osmo has found success in the fragrance industry and won a grant for developing insect repellents. But this will not transfer to food, because there are many molecules, not one. Once again we encounter matrix effects, the thing that makes food science so hard: every ingredient is highly nonlinearly dependent on every other ingredient, as well as the food’s texture, fat content, and many other factors. Mix two flavors, and the resultant combined flavor will not be a neat combination of the two. Put the same flavor in different products, and they will taste completely different: try putting fish sauce in a beer vs a Thai curry. Other cognitive processes also completely dominate our perception. Label the same scent as either vomit or parmesan, and people hate it or love it respectively. Add orange dye to cheddar, and it tastes more intense, whether you’re a regular person or a trained panelist. Serve coffee from a white mug, and it will taste more bitter than if served in a clear glass. Then consider that Korean people like Korean food much more than Indian people like it. Habituation, culture, and memes play a huge role in flavor perception. This is not digitizable. At best, tools like Osmo will help ideate for molecules to put in our food.

Even the codification of sensory science into a discipline hasn’t really worked for making tasty new foods. Highly trained people are great at picking out different descriptors, but this is not the same skill as making good products. Instead, the sensory science fields and the equipment merely excel at making products consistent. Most of the job of a food manufacturer is to keep things the same over decades of supply chain turnover and batch-to-batch variation. A team of sensory testers tease out the minute differences between every iteration, and can force the Texture Profile Analysis voodoo to be the same each time. But none of this helps with making good new products.

Reality check

In practice, a food scientist makes something, everyone eats it. Someone with buy-in says “damn, that slaps”, and they give it to a panel of random people to try and see if they like it, then they ship it. If sensory science or computational smell or taste was getting anywhere, we would have seen it by now. Have you tasted vegan cheese? It’s terrible. Good foods are not done by trying to codify or predict flavor. You tinker around, invent Diet Dr. Pepper, it does well and survives in the market, and you spend decades making sure the flavor never changes.

Just think for a second what it would mean to build a way to get a machine to taste things. We would need to recreate every receptor for smell, taste, touch, sight, and sound, including the hundreds of different odor receptors, and capture the full range of the intricacies of mouthfeel. But that’s the easy part! The hard part would be solving cognition by building a digital cerebellum to figure out how to piece all these signals together. This is insane.

But there is good news. We do not need to build a digital mouth to approximate how people might like a food. We can just give it to them, and ask.

The true way

We will not be able to get AI to taste things. That’s ok. Let’s treat it like the iterative empirical problem that it is. Let’s make a lot of stuff and give it to people to eat, and see if they like it. Let’s do this over and over and over again.

The self-driving food science factory of the future churns out thousands and thousands of products a day. Lines of tens of thousands stretch into the distance, each holding their tasting spoon ready for their Universal Basic Ingestion. Their work has been displaced by AI but they are hungry to advance the great project. They hold out their spoons for the day’s experimental marvels, each more delicious than the last. Their glee radiates from their faces and is logged by the machines. I think we will have jobs in the future after all.

Alternative proteins - plant-based foods, precision fermentation, and lab-grown (cultivated) meat - are an incredibly hard set of problems in the life sciences. There has been tremendous progress in the last decade, products and technologies are far better, but we are still far from having a steak without the animal. This is also one of the world’s most important problems. When we think about people hundreds of years from now looking back at us doing factory farming with disgust, we think one of the main ways we get there is through lab-grown food. There is a world where AI solves lab-grown meat. AI systems will be doing science in all fields, starting with speeding up researchers and, in the not too distant future, automating and roboticizing huge amounts of their work. Efforts are currently focusing around the life sciences. This makes me really excited. We will be able to vastly speed up alternative protein research.

What I want to stress is that the transformational benefits will come downstream from being beneficiaries of generalizable AI, and not from the alternative protein field attempting to build any domain-specific foundation models. The industry has, for a while, thought that many problems could be solved by something that looks like the prophesized AlphaMeat, a model specific to alternative protein or specific areas within it, built by combining data from companies and projects across the industry. I used to think this was possible, and can show you my pitch decks, but I now think we should abandon any efforts around making domain-specific foundation models, and instead redirect our efforts towards the wholesale rebuilding of the industry around automated research and AI Scientists. I think this is probably true for many areas of science.

Something fairly amazing happens when you combine a simply stated legible problem, tons of data, and deep neural nets. Through the input-output prediction, the deep layers gain understanding of the structure of the problem. Alphafold used decades of collected data on sequence-structure pairs to predict protein structure, and its layers learned how proteins fold and how their amino acids interact. Computer vision models, through hundreds of thousands of examples of labeled images, learned spatial representations and structures, which we know because we can use transfer learning to reuse the embedding layers for new classification problems. Then LLMs, through the simply stated goal of predicting the next token, have caused the emergence of intelligence deep in the layers.

There is no clean problem, and definitely not enough data to do this in alternative proteins. The problem we’re facing is to find the right ingredients to cram together in the right way in order to make something so good and so cheap that it will take animal products off the shelf. I genuinely believe this might be one of the hardest problems in the life sciences, because of how non-legible it is. It’s really not clear what specific problem we want to solve, let alone clean input-output pairings. Think about cheese. Cheese is incredible and difficult because of the fascinating protein casein, which is responsible for the coagulation of milk, along with stretching, melting, and browning. How this usually works when we make plant-based cheese is that someone picks some proteins, starches, and fats, and tries different combinations of them with various blending and heating parameters, then adds various premade flavors in the right ratios until you get something that a taster likes. There is no problem here that can be described as a neat input-output machine learning problem. You can’t just say ingredients are the features and a taste vector is the label, because there are infinite ingredients, and we can’t figure out a reasonable tokenization or featurization of ingredients or molecules that are useful here. And you definitely don’t have tens of thousands of data points to learn from, you get a few dozen if you’re lucky, and then you pick a different ingredient set that is a completely different domain to what you’ve seen before. Even changing one ingredient means you usually have to start your data collection all over again, because it causes all kinds of nonlinear changes. There are infinite combinations of ingredients and processes, and the thing we’re trying to get right is how thousands of molecules and the structure they’re delivered in interacts with all the receptors in our mouth and nose, along with our sight, touch, and hearing. In addition, we have not been able to find any machines that are reasonable proxies for taste, we cannot improve on just giving a product to a person and asking questions, which means that research cannot be miniaturized.

One response you might have is, instead of going for good taste, to rather go for identical biomimicry of the product, which would be a clearer problem to measure progress on. This is the approach used in precision fermentation (creating animal molecules) and lab-grown meat (creating animal cells). In the case of cheese, we develop yeast that can produce casein directly, real animal protein made in a bioreactor by yeast. The product is so real that it causes the same allergies and needs to be labeled as an animal product. But this still doesn’t solve the problem of making cheese. It turns out even had we started with the real thing, casein isolate from real milk, no one has so far been able to reconstitute casein isolate into cheddar. We have not been able to break milk down into its constituent parts and rebuild it. It seems that process is as fundamental as ingredients. We won’t just have steak by having all the cells that make up a cow. This means we still have the highly non-legible problem of how to combine things together in a way that tastes good.

There is no simple, legible problem in alternative proteins. There is no escaping the fact that we just have to roll up our sleeves and muddy through iterative empirical science. There is no foundation model coming to save us. It won’t work, because we do not have a problem to solve. We have thousands of different problems, all distinct from each other. There is no point in combining data sets from all over the industry. Even if it was good data, which it generally is not, then we would not get anything out of it apart from understanding the particular problem that a particular lab was trying to solve, that is not useful to anyone else. Every cell line is different. Every piece of manufacturing equipment is different. The optimizations that companies do on their cells are completely contextual to the specific environment they’re doing it in. That is fantastic, that is the work that needs to be done. But it is not transferable, because everyone is off optimizing in their own little part of the infinite dimensional search space. If learnings were transferable, we would just wait for pharma to solve biomanufacturing then copy paste things over. But that isn’t possible. The same is true on every level of the problem. Once you have cells and want to make it into a product, you need to solve that problem that’s completely specific to your cells. You need to respect that we are optimizing over an infinite-dimensional search space of all the possible inclusions of ingredients and processes, and we are never doing anything truly general.

Just imagine for a second we were able to build a general model that solves lab-grown meat. This would mean that we have some sort of legible model that somehow predicts the meatiness of something from constituent ingredients and parameters. We would fully predict how slight changes in the molecules involved would alter someone’s experience of taste. Think of the representational power that this model would have: understanding the highly non-linear interactions between thousands of compounds including the bioreceptors that make up the human sensory systems. This would essentially be the building of a world model: a deep representation of how physical reality works, rich enough to predict novel outcomes, covering chemistry, material science, and maybe even physics. This is insane. This industry is not going to be the one to collect the data to build this model. The prediction model for meatiness is not going to be THE THING that generates the deep layers that solves chemistry. Maybe someday someone will build a model that does this, but it will not come from this field.

I see basically 2 ways that alternative proteins are solved:

1. AI Scientists and automation vastly speed up finding specific solutions on the critical path to specific problems like lab-grown chicken.

2. A model that completely captures chemistry, physics, and material sciences emerges from work elsewhere. It probably looks something like molecular simulation that people figure out how to scale. This is then used to simulate and predict what works.

I am excited because I think AI Scientists will be able to solve the problems. It’s a hard problem, but not THAT hard of a problem that we need complete molecular simulation. We need to pivot the industry to start building for AI Scientists and automation. It is knocking at our door. In a future post, I will outline the bottlenecks we need to address to get this train moving. Choo choo.

You need to move to San Francisco. Whether you are currently working in the movement, or are thinking about getting into the movement, you need to be here. I don’t care if you run a large organization, I don’t think there’s a good reason not to pick up your life and move here unless you literally cannot. This is the most important moment in history if we are going to stop putting 100s of billions of animals through factory farms, the closest approximation to hell.

Power is concentrating. You don’t need to have crazy predictions about AI to see that. San Francisco is quickly becoming the capital of the world. This is a small city of 800,000 people where billions of dollars are flowing every day. There is a narrow period here where we can influence the people and culture of this city to get them to care, but we are missing that boat. The companies here are gobbling up more and more industries, and this would be true even if AI didn’t keep getting better. But AI systems are getting better, and things are about to get monumentally more crazy and unequal.

Whether you like it or not, the people in this city are the people who decide how that future goes, what gets cared about, and where resources get spent. If they want to, they will end factory farming. We are a paltry (lol), small movement with about 2,000 people and only about $300M a year across the whole world. Fortunately for us, this city has thrown many, many billions at things before, and they’ll do it again if we can convince them that what we are doing is cool, valuable, and effective, and because we are their friends and they like us. Because of the effect of EA on the city, animal welfare is already in the water. But it’s not enough, and that’s why you need to get here and push it further. Do not underestimate the effect you can have on the culture by just being here, being interesting, working on interesting things, and being friends with people. The world moves from the house parties in this city.

We are critically underinvested in San Francisco. There is a small and mighty rag tag crew here, trying to get people to care, but we need it to be so much bigger. And we need projects and organizations here. Sentient Futures has quickly become the most important community building organization in the world, but we need far more than that.

So what would it look like to do this properly? It means huge chunks of the movement moves here. I look at the AI Safety movement which has gone from a few scattered researchers to a behemoth in a few years, with many designated buildings around the Bay. This has vastly improved the communication and speed of their ecosystem. They have centralized to do their work, run fellowships, deploy grants, and make a huge influence on the city and the future. We can do this too. We get our organizations to set up their bases here and move people over. We set up exciting new projects, fellowships, and incubators. We become friends with everyone, and turn San Francisco into the city whose legacy includes ending factory farming. This will take time. But it needs to start with specifically you moving here to do what you can to get it started.

It was hard for me to move to San Francisco. I dreaded it. I am from Australia, the best place in the world. I love culture, music and wine, and having friends and communities that I’ve known for decades. The tech scene in SF is not known for these things. And it is hard to be here as an immigrant. But it is where the future is happening. And it is where you will find people, who, like yourself, are willing to acknowledge that and have come here to make it go better. And frankly, because it lacks culture, you can have a big impact on it.

I don’t have a specific job for you. You’ll need some grit to figure out what the plan is. But there is so much to do, and as long as you are motivated we can make it work; helping you get cheap accommodation, introducing you to people, starting you on project ideas, and even training you into an AI Safety researcher so you can make a dent on how AI gets developed in a way that is better for animals.

You have probably committed your life or made a lot of sacrifices already to be in the animal movement. You have done it before and I am asking you to do it again. Moving to San Francisco is probably the best thing for animal welfare the movement you will ever do.

Scott Alexander, one of SFs most important writers and intellectuals, wrote in You Have Only X Years To Escape Permanent Moon Ownership to criticize the boring, wealth-hoarding ambitions of many in SF who want to “escape the permanent underclass”:

In such a situation, greatness is pathetically easy. A random woman gave Jesus a washcloth to wipe his face on the way to his crucifixion. She is now known as St. Veronica, patroness of laundry workers, and one out of every 2,500 girls in America is named in her honor. She has an annual feast day, approximately one million beautiful Renaissance paintings, a chapel in Jerusalem, and lesser churches all around the world (including one here in San Francisco). The richest olive merchant in Jerusalem that year is long forgotten, but she endures.

I hope this resonates with you. We all aim to be Veronica, we do the right thing just because it is right, and hope our deeds pass through the ages. But I also hope you also see the tremendous importance of being in the right place—she was, very importantly—in Jerusalem.

Get here!

PS: I will admit, I’m actually being narrow. I think there’s other places you’re allowed to be. San Francisco, Washington DC, or China. Maybe India or Brazil. End of list. If you’re working in those other places it might be ok to stay there.